Me Learning LLM Reasoning
-
By
happenstanceYouTube algorithm, I stumbled upon these two related videos:Video #1 — a respected AI researcher’s lecture given to Standard engineering students (in English):
Link to videoVideo #2 — a YouTuber’s retelling of that lecture (in Mandarin):
Link to videoI noticed Video #2 first, but realized very quickly that Video #2 is a retelling of Video #1, so I figured why view Video #2 when I can just view the original Video #1.
I very quickly got my answer: as good a scholar as the presenter of Video #1 is, his lecture is quite hard to follow. The pacing, the storytelling, the command of the language, the structure of the presentation, the presentation, etc. are not very good. A brilliant researcher, but not a very good lecturer. I
gave updecided to try the alternative 10~15 minutes into Video #1.Video #2 is 1/3 the length of Video #1, but overall presentation is much nicer. Some terms of art I have to mentally translate back to English to map them back to what I am more familiar with, but with key slides being shown in the original English, that’s not too hard to do. Watched through Video #2 at 1.5x speed and got what I wanted in maybe 1/5 the time I would have spent on Video #1 had there not been Video #2.
There probably is a lesson somewhere in this experience for multilingual LLM machine learning/reasoning, but I don’t know what that is yet.
-
By
happenstanceYouTube algorithm, I stumbled upon these two related videos:Video #1 — a respected AI researcher’s lecture given to Standard engineering students (in English):
Link to videoVideo #2 — a YouTuber’s retelling of that lecture (in Mandarin):
Link to videoI noticed Video #2 first, but realized very quickly that Video #2 is a retelling of Video #1, so I figured why view Video #2 when I can just view the original Video #1.
I very quickly got my answer: as good a scholar as the presenter of Video #1 is, his lecture is quite hard to follow. The pacing, the storytelling, the command of the language, the structure of the presentation, the presentation, etc. are not very good. A brilliant researcher, but not a very good lecturer. I
gave updecided to try the alternative 10~15 minutes into Video #1.Video #2 is 1/3 the length of Video #1, but overall presentation is much nicer. Some terms of art I have to mentally translate back to English to map them back to what I am more familiar with, but with key slides being shown in the original English, that’s not too hard to do. Watched through Video #2 at 1.5x speed and got what I wanted in maybe 1/5 the time I would have spent on Video #1 had there not been Video #2.
There probably is a lesson somewhere in this experience for multilingual LLM machine learning/reasoning, but I don’t know what that is yet.
@Axtremus said in Me Learning LLM Reasoning:
There probably is a lesson somewhere in this experience for multilingual LLM machine learning/reasoning, but I don’t know what that is yet.
Well, to me one obvious lesson is not about machine learning, but rather that, even if you’re very knowledgeable, you have relevant information, and your ideas are good, how you present it matters immensely. Being able to present complex information in an accessible way is the essence of good teaching.
I notice that video #1 says
High-level overview of reasoning in large language models, focusing on motivations, core ideas, and current limitations. No prior background is required.
But reading your comment makes me wonder if in fact prior background is needed… Because being able to present new ideas in an accessible to someone with no background in the subject is the essence of good presenting.
I’m curious about the subject matter in the video (for example, I want to hear about chain-of-thought prompting and chain-of-thought reasoning, and I want to know what they mean by in-context learning), but reading your post makes me hesitant to watch it…
